Creating smart chatbots – an example for banking


This article is published in collaboration with our partner The Campfire, expert in the field of conversation design and chatbot implementation. Their lean approach aims to build chatbot solutions that are engaging, scalable & integrated. Together, we built smart chatbots for international clients in the banking industry, like Belfius and Rawbank.

A lot has already been written about best practices for chatbots. Although many topics might still be controversial, there is one major statement on which everyone agrees: creating a chatbot is not an exact science.

Each project requires a unique perspective and – as so many things in life – better be safe than sorry! Don’t jump straight into building a conversation, because you will end up with bugs, chaos and frustration.

We at created an intuitive platform accessible to every user. However, even the most expert chatbot builders need to think first about what they want to achieve with their bot. Take the example of a big institution, like a bank. When building a chatbot it is key to identify:

  • Which clients do we want to reach?
  • On which channels do we want to include our chatbot?
  • What is the final goal of every chatbot conversation?

For a bank, there might be multiple answers to these questions. Take the example of RawBank, the largest bank in Congo, with a history of over 20 years providing financial services to local businesses. The Campfire helped them building a chatbot on their website, powered by the technology.

Its great UX allows web visitors to easily collect information about the different services offered by the bank, without manually browsing for a specific page.

So, how does a banking bot project begin?

When our partner The Campfire starts a chatbot project, they collect the right answers from main stakeholders that have experience in talking to their target audience. And who is in a better position for this qualification than people from customer support?


To ensure the best result, The Campfire maps out the most frequently asked questions (FAQs) from customer support. In each project, anything between 60 and 80% of requests can be grouped into 20 to 30 questions. Indeed, people often have the same issues or concerns.

There is no use in focusing on fringe cases and creating complex flows for a niche use. This can be done at a later stage of the project. So, when starting to build a chatbot, we should focus on categorizing around 20 to 30 of the most important topics. Thanks to the input from customer support, The Campfire team can define the questions and start grouping them together.

Grouping is important! Within the Chatlayer platform, it’s easy to build visual conversation flows. And because we think of grouping from the very beginning, we can create what we call “smart fallbacks”. But this is best explained with an example.

Flow grouping

A lot of questions to the customer support team are about pricing. The general intent of the user is “I want to know the price of…”. This intent can be reused for different products. The user might want to know the price of a VISA card or of a specific savings account. By combining these questions in the same intent, we make sure that the Natural Language Processing (NLP) model is trained in the best possible way. Even with our strong NLP model, it is important to think about grouping relevant expressions together to make the bot more accurate in understanding what users are looking for.

What is an intent? Read more about it here!

If the chatbot now recognizes “VISA card”, it is able to say:
“What would you like to know about our VISA cards?”        
and show the available options for which an answer is prepared. Make sure to always add the option “I have another question” and provide an appropriate transfer to the customer support.

Using this logic, The Campfire creates chatbots that are able to answer almost any question a user might ask:

  • If the chatbot recognises the full intent and entity (eg: “I want to know the price of a VISA card”), there is a complete answer available in
  • If the chatbot recognises only the intent but not the entity (eg: “I want to know the price of everything”), we have a clear fallback. The chatbot will reply “For which product would you like to know the price?” and show the available options to the user.
  • If the chatbot recognises only the entity but not the intent (eg: “my VISA card is green and I want an orange one”), we have another clear fallback. The chatbot will reply “What would you like to know about our VISA card” and show the available options to the user.
  • If the chatbot does not recognise any intent or entity (eg: “my girlfriend is getting on my nerves”), we have a final fallback. The chatbot will reply “I am not sure I can help you with that. I can help you with the following topics:” and show the available topics.
This main structure allows The Campfire to build a smart chatbot for large companies, capable of capturing as much information from the conversation as possible. This model has proven to be extremely scalable and results in very high NLP recognition. It’s easy to add more questions to the flow in as you’ll move forward with your chatbot project.

Once your flow is ready, it’s time to look at the best channels to reach your target audience. 

Do you want to know more about the next steps?

Contact us and get the right tips from our partner The Campfire on how to build better bots!

A collaboration between: 

Want to know more?